BACKGROUND: A Bayes Network was developed for individual risk prediction after cholecystectomy. Validity and robustness were compared with logistic regression analysis (LR). METHODS: Clinical databases were created at the Ulm University and St. Franziskus Flensburg hospitals between 2001 and 2010 were comprised of hospitalized cholecystolithiasis patients serving as model and test cohorts, respectively. The probabilities of in-hospital death, prolonged hospitalization (>7 days), relaparotomy and erythrocyte transfusions were predicted based solely on admission data by BN and LR. ROC curves were calculated. RESULTS: The Ulm and Flensburg cohorts consisted of 1,029 and 1,842 patients, respectively. The areas under the ROC curves for predicting death were 94% (p = 0.8) for both BN and LR, 70 vs. 76% (p < 0.001) for prolonged hospitalization, 69 vs. 68% (p = 0.8) for relaparotomy, and 84 vs. 78% (p = 0.1) for ET. Predictability declined for both methods when explanatory values were changed randomly. In contrast to LR, the BN revealed a good robustness to missing values. CONCLUSION: Both BN and MR predicted the death risk quite accurately. The advantage of BN consists of its robustness to missing values. Moreover, its graphical representation may be helpful for clinical decision making.
BACKGROUND: A Bayes Network was developed for individual risk prediction after cholecystectomy. Validity and robustness were compared with logistic regression analysis (LR). METHODS: Clinical databases were created at the Ulm University and St. Franziskus Flensburg hospitals between 2001 and 2010 were comprised of hospitalized cholecystolithiasispatients serving as model and test cohorts, respectively. The probabilities of in-hospital death, prolonged hospitalization (>7 days), relaparotomy and erythrocyte transfusions were predicted based solely on admission data by BN and LR. ROC curves were calculated. RESULTS: The Ulm and Flensburg cohorts consisted of 1,029 and 1,842 patients, respectively. The areas under the ROC curves for predicting death were 94% (p = 0.8) for both BN and LR, 70 vs. 76% (p < 0.001) for prolonged hospitalization, 69 vs. 68% (p = 0.8) for relaparotomy, and 84 vs. 78% (p = 0.1) for ET. Predictability declined for both methods when explanatory values were changed randomly. In contrast to LR, the BN revealed a good robustness to missing values. CONCLUSION: Both BN and MR predicted the death risk quite accurately. The advantage of BN consists of its robustness to missing values. Moreover, its graphical representation may be helpful for clinical decision making.